NateRunsA-LIST commited on
Commit
8f2caee
·
verified ·
1 Parent(s): 94d3c60

README: v2 sol-refit release notes

Browse files
Files changed (1) hide show
  1. README.md +16 -6
README.md CHANGED
@@ -14,17 +14,27 @@ library_name: coreml
14
  On-device replacement for the remote LLM that extracts structured numbers from
15
  free-text exercise entries in the A-LIST app (locomo mode). Single fp16 Core ML
16
  graph: MiniLM-L6 encoder (frozen, mean-pool + L2-norm inside the graph) + MLP
17
- heads, distilled from ~14.5k historical production extractions.
18
 
19
  - **Inputs**: `input_ids` int32 (1,128) WordPiece/bert-base-uncased; `attention_mask`
20
  int32 (1,128); `parsed` float32 (1,10) regex features (see the app's
21
  `ExerciseFeatureParser` for the exact order and transforms).
22
  - **Output**: float32 (1,8) — `[log1p(duration_min), log1p(calories), heart_rate_bpm,
23
  intensity_1_10, steps_gate_logit, log1p(steps), distance_gate_logit, log1p(distance_km)]`.
24
- - **Eval** (2,795 held-out-user entries, agreement with the production LLM):
25
- duration 77% (±20% or 3 min), calories 71% (±25% or 20 cal), heart rate 87%
26
- (±15 bpm), intensity 83% (±1), steps/distance zero-gates 92–94%.
27
- - `golden_cases.json` holds 10 input→output cases generated from this exact
28
- artifact for integration testing.
 
 
 
 
 
 
 
 
 
 
29
 
30
  Internal model for the A-LIST iOS app. Not intended for third-party use.
 
14
  On-device replacement for the remote LLM that extracts structured numbers from
15
  free-text exercise entries in the A-LIST app (locomo mode). Single fp16 Core ML
16
  graph: MiniLM-L6 encoder (frozen, mean-pool + L2-norm inside the graph) + MLP
17
+ heads.
18
 
19
  - **Inputs**: `input_ids` int32 (1,128) WordPiece/bert-base-uncased; `attention_mask`
20
  int32 (1,128); `parsed` float32 (1,10) regex features (see the app's
21
  `ExerciseFeatureParser` for the exact order and transforms).
22
  - **Output**: float32 (1,8) — `[log1p(duration_min), log1p(calories), heart_rate_bpm,
23
  intensity_1_10, steps_gate_logit, log1p(steps), distance_gate_logit, log1p(distance_km)]`.
24
+
25
+ ## Versions
26
+
27
+ - **`v2/` (current, 2026-08-23)**: identical graph retrained on the sol-relabeled
28
+ dataset (exercise-text-unified:v0 — one uniform frontier teacher, calorie CV 1.6%
29
+ vs the old labels' ~11%). On the frozen held-out-user sol eval (1,127 entries) it
30
+ beats the v1 artifact on every field: duration 68.5% vs 62.7%, calories 58.0% vs
31
+ 51.5%, heart rate ±10bpm 76.0% vs 67.6%, intensity ±1 81.7% vs 76.0%, steps gate
32
+ 94.9% vs 88.6%, distance gate 94.6% vs 90.8%. Same I/O contract and footprint.
33
+ - **Root files (v1, 2026-08-16)**: distilled from ~14.5k historical production
34
+ extractions (mixed teacher generations). Kept resolvable for rollback; the app's
35
+ `ExerciseExtractorDescriptor.v1()` still pins these bytes.
36
+ - `golden_cases.json` holds 10 input→output cases generated from the v1 artifact —
37
+ the app uses them to pin the *decode rules* (weights-independent), so they remain
38
+ valid across versions.
39
 
40
  Internal model for the A-LIST iOS app. Not intended for third-party use.